1

A corpus-based, pilot study of lexical stress

variation in American English

Abstract

Phonological free variation describes the phenomenon of there being

more than one pronunciation for a word without any change in meaning

(e.g. because, schedule, vehicle). The term also applies to words that

exhibit different stress patterns (e.g. academic, resources, comparable)

with no change in meaning or grammatical category.

A corpus-based analysis of lexical stress variation is one way of testing

the validity of surveys of speakers‟ pronunciation preferences for certain

variants. Such surveys include Wells‟ surveys of British English (1999,

2008) and Shitara‟s survey of American English (1993). The present paper

presents the results of a pilot study of American English, replicating part

of Mompéan‟s corpus-based study of British English (2010).

In the current paper, the corpus consists of talks from the TED website

(http://www.ted.com), covering the period February 2002 to June 2009.

The corpus includes approximately 11.5 hours of transcribed speech

(92,750 words) produced by 34 educated speakers (17 men and 17

women with an American accent of standardized variety and some traces

of regional pronunciations).

To guarantee a minimum of representativeness for this pilot study, the

items analyzed were found at least ten times in the corpus. The preliminary

list of search terms showing lexical stress variation was based on the 261

items in the 2008 Longman Pronunciation Dictionary for which survey

data was provided, lists found in a textbook on American pronunciation

(Celce-Murcia et al., 1997/2007), lists in two previous studies (Shitara,

1993: Mompéan, 2010) and from anecdotal knowledge of frequent

variants. Detailed results for lexical stress are provided for seven items.

The TED corpus results do not always concur with LPD data and raise

interesting issues concerning the use of authentic spoken corpus data. The

paper also discusses designing and carrying out corpus-based

pronunciation studies.

hal-0

0636

624,

ver

sion

1 -

27 O

ct 2

011

Author manuscript, published in "Research in Language 8, 8 (2010) 1-15"

2

I. Introduction

In general, phonological free variation describes the phenomenon of

there being more than one pronunciation for a word without any change in

meaning (e.g. because, schedule, vehicle). The term also applies to words

that exhibit different stress patterns (e.g. academic, resources,

comparable) with no change in meaning or grammatical category.

According to Mompéan, phonological free variation may occur for a

variety of reasons, which may interact: sound change, phonetic processes

and cognitive or sociolinguistic/sociocultural factors, where analogy

might affect lexical stress (2010). In his 2010 study, he excluded

homographs and variation due to changes in grammatical category,

which is entirely justifiable in a study of phonemic variation. However,

in his study variation also had to be a characteristic of citation forms and

therefore he excluded variation due to rhythmic, contextual influences.

Applying the last criterion to a study of lexical stress variation would

make it extremely difficult to find enough occurrences in naturally

occurring speech but, more importantly, would exclude from analysis a

potentially rich locus of variation.

One source of lexical stress variation due to rhymthic, contextual

influence is that of stress shift. In Cruttenden, accentual variation confirms

“the tendency in English to avoid adjacent accented syllables. It is in order

to avoid the placing of primary accents on adjacent syllables that „accent

shift‟ occurs in phrases such as ‘Chinese ‘restaurant (but Chi’nese) ...”

(2001, 280). Rhythmic constraints can be among the most difficult for

teachers to explain and for learners to acquire; it is therefore essential

that they be addressed in any publication that seeks to prioritise

competing pronunciations. This pedagogical reality emphasizes the need

for more studies that use corpora to verify preference poll data.

This paper is a corpus-based pilot study of lexical stress variation in a

corpus of modern American English. It provides a useful approach for

checking the validity of surveys of pronunciation preferences referred to in

the 2008 edition of the Longman Pronunciation Dictionary (LPD), and

which are meant to provide “some kind of objective data regarding the

relative prevalence of competing pronunciations of various words”

(Wells, 2003, 215). As such this paper tries to replicate a small part of

hal-0

0636

624,

ver

sion

1 -

27 O

ct 2

011

3

Mompéan‟s 2010 broader study of free variation in British English, in

relation to the LPD (2008) surveysi.

II. Method

II.1. Data: Corpus Creation

In order to create a spoken corpus for further study, various on-line

sources were explored. As variation in stress involves connected speech

processes, dictionary sites and CDs were excluded because they typically

provide citation-form pronunciationsii.

A major criterion in the corpus design was that a transcription must

accompany the sound file, in order to eliminate time-consuming

transcribing work. However, other factors also influenced the choice of

sources. For example, the Voice of America covers current affairs on its

ESL/EFL site and provides transcriptions to accompany sound files which

can be downloaded. A variety of American voices are used and the

majority have long stretches of monologue speech. However, as these are

designed for learners they tend to involve slow, careful pronunciation that

cannot be seen as representative of normal, everyday connected speech.

Similarly, the NOVA ScienceNow site also looked like a promising source

of podcasts, as transcripts were readily available for free. Unfortunately,

these tend to involve several speakers.

The final choice for this study was sound files from videos on the TED

website. TED is a small nonprofit organisation in the United States

“devoted to Ideas Worth Spreading”. It started out in 1984 as an annual

conference bringing together people from Technology, Entertainment and

Design. Videos of these talks are stored on-line, along with interactive text

transcriptions and subtitles in various languages provided by viewers (see

Figure 1):

hal-0

0636

624,

ver

sion

1 -

27 O

ct 2

011

4

Figure 1. Screenshot of www.ted.com

The talks selected for the present study cover a variety of topics from

February 2002 to June 2009. The talks are listed in Appendix A. The

transcriptions range from 275 to 5150 words in length. The corpus

includes approximately 11.5 hours of transcribed speech (92,750

words) produced by 34 speakers (17 men and 17 women) with an

American English accent. As this study looks at variation over a range of

American accents, the corpus was not limited to speakers of a Network

Standard or other “standardized” form. However, given the formal, public

context the selected talks are assumed to represent intelligible, educated

American English, though perhaps exhibiting certain regionalisms. The

speakers range in age from early 30s to early 60s and come from a host of

professions. Further socio-cultural details could be found on-line, as the

identity of all of the speakers is clear.

In order to extract high quality sound files from the videos, AudaCity

freeware (version 1.2.6 Stable) was used. A cable simply joined the

“headphone” output to the “microphone” input and “line input” was

selected as the sound source in AudaCity.

Corpora come in a variety of sizes, subject to both the nature of the

research question and logistical concerns (McEnery, Xiao & Tono, 2006,

72-73). McEnery and Wilson argue that “…the size of the corpus needed

hal-0

0636

624,

ver

sion

1 -

27 O

ct 2

011

5

to explore a research question is dependent on the frequency and

distribution of the linguistic features under consideration in that corpus”

(2001, 80). Larger corpora are needed for studies of lexis than for

grammar, for example, because the validity of conclusions is largely

dependent upon the frequency of occurrence of a word. Research which

seeks to determine which pronunciation variants are most likely to occur

arguably require similarly sizeable corpora, as frequency of occurrence is

the determining factor in ranking variants. At first glance, the size of the

TED corpus is respectable, being intermediate in size between the SEC

and the WSC corpora of spoken English (Table 1):

Name of corpus Size Other

Information

TED Corpus of

American speech

92,750

words

spoken, prepared

monologues

BNC (British National

Corpus)

10 million

words

*spoken = 10% of

total 100 million

words

ANC (American

National Corpus)

22-100

million words

since 1990, in

development

MICASE (Michigan

Corpus of Academic

Spoken English)

1.7 million

words

university speech,

through 2002

LLC (London-Lund

Corpus)

250,000

words

UK, 1960-70s,

monologues

SEC (Lancaster/IBM

Spoken English Corpus)

53,000

words

UK, radio

broadcasts, through

1987

CANCODE

(Cambridge/Nottingham

Corpus of Discourse in

English)

5 million

words

UK, interaction,

through 1997

WSC (Wellington

Corpus of Spoken New

Zealand English)

120,000

words

NZ English,

formal speech, out of

a total of 1 million

words, through 1998

ICE (Internat’l Corpus potentially spoken & written,

hal-0

0636

624,

ver

sion

1 -

27 O

ct 2

011

6

of English) 500,000 spoken

words from each

world English

out of a total of 20 x

1m words of each

world English, since

1989

Table 1. Size of different spoken English orpora

However, as the results and analysis show, the small size of the corpus

meant that a significant number of occurrences was not always obtained.

This limited the number and the robustness of conclusions which could be

drawn from the data, as is often the case in a pilot study.

II.2. Data: Search Terms

Corpus queries are often based on pre-established lists; for this study

the goal is to see how these descriptive lists of pronunciation variants

compare to authentic, connected speech. Such lists are often based on

items found in previous research, dictionaries or textbooks. Using five

such sources, a preliminary list of almost 400 potential search terms was

compiled:

52 items from Mompéan (2010),

261 items in the 2008 LPD for which survey data was provided

and where variable stress would be expected,

37 items from Shitara‟s 1993 opinion poll of American word

stress variation,

numerous items listed in Celce-Murcia et alia‟s textbook on

teaching pronunciation (1997/2007),

9 items from anecdotal/personal knowledge of frequent variants,

eg. development, academic.

None of these sources could be used as the sole search list, because a

preliminary analysis of the TED corpus did not reveal enough occurrences.

In line with Mompéan (2010), items were only included in the study if

they occurred ten or more times in the corpus, giving a final list of eight

items:

hal-0

0636

624,

ver

sion

1 -

27 O

ct 2

011

7

2 items from Mompéan (2010): complex (12 occurrences),

economic (20)

3 items from the LPD (2008): Chinese (61), individual

(17)/individuals (12), Japanese (11).

1 items from Shitara (1993): create (44)

1 item from Celce-Murcia (1997/2007): necessarily (12)

1 item from anecdotal/personal knowledge: research (12)

Items were included even if fewer than ten speakers produced them.

This is a major drawback of data that is not produced in a controlled,

laboratory setting; it is not always possible to collect enough occurrences

of lexical items, nor is it always feasible to control for intra-speaker

variation by getting enough occurrences from different speakers. Table 2

contains the final list of words studied:

Chinese

create

individual

research

complex

economic

Japanese

necessaril

y

Table 2. Lexical items studied in the TED 2002-2009 corpus

II.3. Speakers

Thirty-four speakers were chosen: 17 females, 17 males. Their accents

were classified as American, based on features such as the presence or

absence of rhoticity and typical segmental inventories described for

General American English; six native speakers of English were also asked

to confirm whether or not speakers were native or non-native speakers of

American English, regardless of regional accent. American English is

defined as in the LPD as the accent spoken by most Americans “…..who

do not have a noticeable eastern or southern accent” (LPD, 2008, xx). One

Canadian speaker, Steven Pinker, was excluded because of his nationality

and his accent, which is a mixture of Canadian and GAE features.

hal-0

0636

624,

ver

sion

1 -

27 O

ct 2

011

8

II.4. Procedure

The interactive transcription of each talk was copied into an Excel file

which included: a) the speaker‟s name and background; b) the URL where

the audio file is available; c) the title of the speaker‟s talk; d) the length in

minutes/seconds of the talk; e) the number of words of the talk; and f) the

dates the talk was “performed” and posted. Each sound file was

downloaded and then carefully listened to in order to correct mistakes in

the transcriptions.

Analysis involved four steps: locating the target words in the written

corpus using the freeware concordancer ANTConc (Anthony, 2007);

listening to the relevant sound file on-line; noting each occurrence along

with the speaker‟s name; determining which variant was produced. The

variant was initially identified by the author. When a firm identification

was not possible, items were inspected spectrographically using PRAAT, a

freeware speech analysis tool developed by Boersma and Weenink (2008).

An attempt to use external raters failed, due to faulty editing of sound files

and other design issues. Future research will correct this error.

III. Results & Analysis

For some of the items, several forms (eg. plural, past tense) were found

in the corpus; the term “word family” in Table 3 reflects this reality:

Word Family n° of

Occurrences

n° of

Speakers

CHINESE

61 2

COMPLEX

10 6

CREATE 74 25

create (44)

creates (6)

created (20)

recreate (3)

ECONOMIC 19 7

hal-0

0636

624,

ver

sion

1 -

27 O

ct 2

011

9

INDIVIDUAL 26 12

individual (16)

individuals (12)

JAPANESE 11 4

NECESSARILY 12 9

RESEARCH 19 12

research (12)

researcher (2)

researchers (5)

Total 232 x

Table 3. Number of occurrences in the word list generated by AntConc

Different words in each word family were included in the analysis,

even though only four of them provided ten or more occurrences for ten or

more speakers: CREATE, ECONOMIC, INDIVIDUAL and RESEARCH.

The results for CHINESE were not analysed, as they were skewed by there

being only two speakers.

The analysis of the occurrences shows that the TED speakers do not

always concur with the LPD‟s listed pronunciations for General American

English (GAE), shown in Table 4:

hal-0

0636

624,

ver

sion

1 -

27 O

ct 2

011

10

Table 4. LPD Dictionary pronunciations of the items for RP (Received

Pronunciation) and General American English (GAE).

In Table 4, items in the middle have the same pronunciation in both RP

and GAE. LPD conventions apply:

hal-0

0636

624,

ver

sion

1 -

27 O

ct 2

011

11

italic / / = sound sometimes optionally omitted

raised //sound sometimes optionally inserted

/ / = possible compression of adjacent syllables

= alveolar tap, usually voiced, like in AmEng city

The √ symbol indicates the LPD recommended main pronunciation.

The results for each word family are analysed in more detail in the

following sections. Despite this paper‟s focus on stress variation, one

phonetic process – compression – is mentioned, as it affects the number of

syllables and often lexical stress.

III. 1. Complex

In the adjectival form of the bisyllabic word complex, variation is

commonly expected but the LPD (2008) preferences for American English

showed a distinct preference (73%) for stress on the second syllable.

Interestingly, this pattern was only found twice in the nine adjectival

occurrences, and from two different speakers: incredibly complex and no

matter how complex they are. Table 5 shows the other seven occurrences

from four speakers which are stressed on the first syllable, the opposite of

the variant proposed by the LPD:

Speaker Search item in context

Tulley actually ‘complex things made by other

Tarter find more ‘complex ‘signals

to find faint, ‘complex ‘signals that our

Roach in the ‘complex ‘sensory-motor action

Boston many other ‘complex ‘human motions

conform to the ‘complex ‘topological

shape

to deal with this ‘complex to’pology,

various

Table 5. Occurrences of complex in context, per speaker

All except the last example appear to be cases of stress shift.

hal-0

0636

624,

ver

sion

1 -

27 O

ct 2

011

12

III. 2. Create For the item CREATE, no occurrences of stress shift were found, even

though the LPD lists this as possible. According to the LPD, 87% of

respondents preferred to stress the second syllable; this was the case in

most of the 74 occurrences over 25 speakers. Compression seems to be

occurring in a few cases, so create sounds like crate. This may or may not

be due to regional variation. Unfortunately, at this stage it is impossible to

say precisely how many occurrences concur with the LPD, because the

external raters showed far too much variation in their judgments. Further

studies will examine this in detail, and external raters will be given better

designed stimuli and instructions.

III. 3. Economic

The LPD gives the main pronunciation with stress on <no> and 11 of

the 19 occurrences follow this pattern. Stress shift is not mentioned as a

possibility in GAE and yet six cases were found in one speaker (Pine), as

shown in Table 6:

Search item in context

the predominant ‘economic ‘offering

this progression of ‘economic ‘value

a new level of ‘economic ‘value

becoming the predominant ‘economic

‘offering

are the ‘economic ‘offerings you are

providing

think about the ‘economic ‘value they have

Table 6. Stress-shifted occurrences of economic in context

As no other speakers produced the collocations economic +value or

economic +offering, it is impossible to know whether or not collocational

hal-0

0636

624,

ver

sion

1 -

27 O

ct 2

011

13

factors influence the likelihood of stress shift, but future research could

look into this.

Two other cases of stress shift occurred in two other speakers: caring

about ‘economic ‘factors and is an ‘economic ‘tipping point. Finally,

Table 7 shows four examples which did not exhibit stress shift:

Speaker Search item in context

Carter for environmental and eco’nomic ‘justice

Alcorn Most of the eco’nomic ‘models are built

about ‘social-eco’nomic ‘movements

Rosenda

le

kinds of eco’nomic ‘forces

Table 7. Occurrences of economic in context, per speaker

Whether or not these cases of shift represent speaker-specific

idiosyncracies or regional variations, they are unpredictable cases; the

speakers could have shifted the stress because a word with primary stress

follows.

III. 4. Individual

Stress shift is indicated by the LPD as possible for the word individual,

but none of the 26 occurrences in the present study display this.

Compression, however, was found (regardless of grammatical category) in

6 of the 12 speakers.

III. 5. Japanese

For the item Japanese the LPD lists one pronunciation and stress shift

as possible. In the 11 occurrences in the TED corpus, several examples of

stress shift (regardless of grammatical category) were found over three

speakers (Table 8):

Speaker Search item in context

hal-0

0636

624,

ver

sion

1 -

27 O

ct 2

011

14

Wallace direction that ‘Japanese ‘toilet technology

Baraniuk languages like ‘Chinese, ‘Japanese and Thai

Lee ‘Japanese ‘Chinese food

all the ‘Japanese ‘bakers who introduced

Chinese food and ‘Japanese ‘foods,

the ‘Japanese ‘immigrants came

something that is ‘Japanese to being

locked up all the ‘Japanese during World War

invented by the ’Japanese, popularized

sort of like a ‘Japanese ‘guy coming

Table 8. Stress-shifted occurrences of Japanese in context, per speaker

Wallace‟s shift to word-initial stress is predictable, but not all of the

other examples can be explained by stress clash avoidance: for example,

Baraniuk‟s languages like ‘Chinese, ‘Japanese and Thai but also Lee‟s

something that is ‘Japanese to being something that is ‘Chinese and

invented by the ’Japanese, popularized by the ‘Chinese. The latter two can

be attributed to contrastive stress, as the extended context shows.

However, the speaker could just as easily have maintained initial stress

and expressed contrast.

In one case stress shift actually resulted in stress clash: It was a

Japa’nese scientist who first undertook … It is difficult to ascribe this to

the discursive context. The stress pattern is used in a context where

contrast is not being signalled, as the preceding text is about the vegetation

where Bonobos frequently live:

The wild Bonobo lives in central Africa, in the jungle encircled by the

Congo River. Canopied trees as tall as 40 meters, 130 feet, grow densely

in the area. It was a Japanese scientist who first undertook serious field

studies of the Bonobo, almost three decades ago.

However, a low speech rate may explain this shift. This is scripted

monologue which accompanies a video clip from a documentary film that

the TED speaker showed, so the film-speaker was probably not at a loss

and searching for words.

hal-0

0636

624,

ver

sion

1 -

27 O

ct 2

011

15

III. 6. Necessarily

Speech rate may also explain the compression found in the occurrences

of the word necessarily. The LPD only provides preference data for British

English, finding that 68% prefer primary stress on the third syllable

nece’ssarily and 32% prefer initial stress. This is close to the 25% (3/12)

of occurrences with word-initial stress in the TED corpus of American

English (Table 9):

Speaker Search item in context

Lee who ate rice would’ necessarily bring

down

couldn't ‘necessarily be fraud,

Abrams you wouldn't ‘necessarily think of when

Table 9. Word-initial stressed occurrences of necessarily in context,

per speaker

Analysis with PRAAT showed that compression may have occurred in

two of the very fast speakers, Wallace and Powers. However, external

raters‟ judgments for these two were extremely varied. This not only

reinforces the case for including speech rates in corpus-based studies but

also confirms the well-known difficulty some individuals have in

perceiving syllables and/or stress.

III. 7. Research

The final item, RESEARCH, seems to reflect national and socio-

economic influences. According to the LPD:

the // form appears still to predominate in

universities, although // has increasingly displaced it

in general usage both in Britain and in America. Some speakers

may distinguish between the verb .‟. and the noun „. . . (2008,

683).

hal-0

0636

624,

ver

sion

1 -

27 O

ct 2

011

16

The LPD preference poll of British English found 80% in favour of

word-final stress in the word research, the figure rising to 95% among

university teachers. Conversely, for American English the LPD found a

preference for word-initial stress (78%). Table 10 shows the 4 of the 19

occurrences from the TED corpus that do not have word-initial stress,

including two occurrences of researchers:

Speaker Search item in context

Benyus mainly about re’search in biomimicry.

Tarter generously supported this re’search.

Wallace from some re’searchers at Stanford

that these re’searchers did MRI brain

Table 10. Word-final stressed occurrences of research in context, per

speaker

None of these can be attributed to clash avoidance. Thus, the results in

the TED corpus (15/19 or 79%) confirm the LPD results for American

English.

IV. Conclusion

Phonological free variation, or variation in the pronunciation of a word

without any change in meaning, also applies to words that exhibit different

stress patterns with no change in meaning or grammatical category. Such

variation may occur for several reasons, of which phonetic processes,

sound change and cognitive or sociolinguistic/sociocultural factors.

Which of these variants to prioritise is a recurring problem in modern

dictionaries and the use of pronunciation preference survey polls can be

one solution. However, another solution could be corpus-based studies,

as they can provide greater quantities of more objective data, with a

corresponding increase in the validity of those predictions and perhaps a

reduced “logistical cost”. Consequently, this paper has provided some

initial results from a corpus-based pilot study of spoken American

English, partly a replication of an earlier study of British English by

hal-0

0636

624,

ver

sion

1 -

27 O

ct 2

011

17

Mompeán (2010). Authentic connected speech from the TED corpus was

used to study lexical stress variation, including that due to the rhythmic

phenomenon known as stress shift. Mompéan‟s study of free

phonological variation was much more extensive (2010) but excluded

stress shift; given the current study‟s focus on lexical stress variation the

influence of stress shift was actively sought it out. This proved

productive, as it allowed speech rate and larger discursive context to be

proposed as factors influencing stress variation.

In general, evidence from the TED corpus confirms some but not all

of the preferences in the LPD pronunciation polls. In direct contradiction

to the LPD, seven of the 9 occurrences of complex are stressed on the first

syllable; all seven also seem to be cases of stress shift. The items create

and individual showed compression but no stress shift.

Analysis of the results for two other items, economic and necessarily,

raised the possibility that two other factors might play a role in stress shift.

Eight occurrences of economic showed shift despite it not being mentioned

in the LPD and six of the occurrences from one speaker hint at the possible

influence of collocational knowledge. This speaker used the collocations

„economic „value and „economic „offering; it is not impossible that

collocational factors affect the likelihood of stress shift, and future

research could look into this. Secondly, the LPD does not provide any

data for American preferences for necessarily. Notwithstanding, 25% of

the examples in the TED corpus use word-initial stress, which is not very

different from the LPD‟s finding of 32% for British English. Speech rate

might help to explain the compression found in several occurrences of

necessarily.

For the item research, the results from the TED corpus (79%) confirm

the LPD results (78%) for American English. However, it must be

emphasized that given the small size of the corpus and the under-

representation of several items, none of the statistics can be used to

generalize about lexical stress in American English. Moreover, the LPD

data reflect the preferences of people from various social backgrounds,

which is also true for the TED corpus. TED speakers tend to be well-

spoken, articulate individuals, with varying levels of academic

qualification. This may skew the results for research, where the LPD

poll shows different preferences for university teachers and other

respondents.

hal-0

0636

624,

ver

sion

1 -

27 O

ct 2

011

18

Perhaps the most interesting item is Japanese, as 3 of the 11

occurrences cannot be explained by stress clash avoidance and one shift

creates a clash. The first two occurrences reveal the speaker‟s desire to

contrast two nationalities but the third might result from a low speech rate.

As compression and other connected speech processes are more likely to

occur when the number of unstressed syllables and the overall speech rate

are increased, perhaps stress shift becomes less predictable when speech

rates are lower. An objective measure of speech rate may need to be

included in corpus-based studies of spoken language such as this.

Despite all the insights provided and the data obtained, it should be

borne in mind that the present study is only a pilot study and, as such,

has severe limitations. It was impossible to control the number of

speakers and occurrences, so certain items are under-represented, which

means that no claims can be made about the relative prevalence of free

variants. Age differences were not explored but a large, diachronic

corpus could potentially address this. Likewise, regional differences

were glossed over, assuming that one General American English exists

with shared recognizable tendencies. Finally, verification by external

raters, which can be useful, was not possible and would have to be

integrated into further work on these issues.

Directions for future research include addressing all those issues in

further corpus-based studies or controlled production tasks. Nonetheless,

this pilot study confirms that such corpora can be usefully designed to

verify survey data. English is a living language and preferences are

bound to evolve. This is a major argument in favour of using a large

corpus (which can be easily updated to track diachronic change) in

addition to survey data, in order to decide which pronunciation(s) to

prioritise not only in dictionaries but also in language teaching.

The use of such polls in pronunciation dictionaries gives teachers and

learners access to explanations about current usage. They can then

organise that knowledge into rules which help them to predict lexical

stress patterns. For example, if General American English (GAE) speakers

tend to final-stress bisyllabic French loan words, then ‘garage is probably

a British pronunciation and ga’rage is probably GAE. This ability to

predict gives learners more autonomy, which is the goal of much teaching:

independent application of appropriate knoweldge in new contexts.

Similarly, easy access to digital resources in today‟s world means that

it is no longer adequate to take at face value statements such as “Speakers

hal-0

0636

624,

ver

sion

1 -

27 O

ct 2

011

19

of Canadian English tend to stress the second syllable in words x, y and z.”

Teachers have the ability to collect data for themselves and/or access data

via on-line journal subscriptions, forums, etc. This enables them to analyse

first-hand what is happening, for example, in American English today.

Bibliography

Anthony, L. (2007). ANTConc (Version 3.2.1) [Computer Program].

Retrieved June 2007, from

http://www.antlab.sci.waseda.ac.jp/software.html.

AudaCity. (2009). (Version 1.2) [Software]. Available from

http://www.audacity.sourceforge.net

Boersma, P. & Weenink, D. (2008). Praat: doing phonetics by

computer

(Version 4) [Computer program]. Retrieved January 20, 2008, from

http://www.praat.org/.

Celce-Murcia, M, Brinton, D.M. & Goodwin, J.M. (1997/2007).

Teaching

pronunciation: A reference for teachers of English to speakers of

other languages. Cambridge: Cambridge University Press.

Cruttenden, A. (2001). Gimson’s pronunciation of English (6th

edn).

London: Arnold.

Free Dictionary. [Website]. http://www.thefreedictionary.com/squishy

Accessed June 2009.

McEnery, A. & Wilson, A. (2001). Corpus linguistics (1st edn 1996).

Edinburgh: Edinburgh University Press.

McEnery, A., Xiao, R. & Tono, Y. (2006). Corpus-based language

studies: An advanced resource book. London: Routledge.

hal-0

0636

624,

ver

sion

1 -

27 O

ct 2

011

20

Merriam-Webster Dictionary. [Website]..

http://www.merriam-webster.com/help/audiofaq.htm. Accessed

January 2009-February 2010.

Mompéan, J.A. (2010). A corpus-based study of phonological free

variation in English, in Henderson, A.J. English Pronunciation: Issues

& Practices, Conference proceedings of EPIP I, June 3-5, Université de

Savoie, Chambéry, France. Chambéry, France: Presses de l‟Université de

Savoie. (forthcoming).

NOVA ScienceNow. [Website].

http://www.pbs.org/wgbh/nova/sciencenow/ Accessed March

2009.

Shitara, Y. (1993), “A survey of American pronunciation

preferences”. Speech Hearing and Language 7, 201-232. Available

at www.phon.ucl.ac.uk/home/wells/shitara.pdf

TED. [Website]. http://www.ted.com. Accessed January 2009-

February

2010.

Temperley, D. (2009). Distributional Stress Regularity: A Corpus

Study.

Journal of Psycholinguistic Research, 38, 75-92.

Trudgill, P. & Hannah, J. (2008). International English: A Guide to the

varieties of Standard English. London: Hodder Education.

Voice of America. [Website]. http://www.manythings.org/voa/rss/

Accessed June 2009.

Wells, J.C. (2008). Longman Pronunciation Dictionary. 3rd

edition,

Harlow, England: Pearson-Longman.

Wells, J.C. (1999). “British English pronunciation preferences: a

changing

hal-0

0636

624,

ver

sion

1 -

27 O

ct 2

011

21

scene”. Journal of the International Phonetic Association 29 (1), 33-

50.

Appendix A

TED corpus: List of speakers and URL

Speaker's Name URL : http://www.ted.com/talks

Pete Alcorn

pete_alcorn_s_vision_of_a_better_w

orld.html

Benjamin Wallace

benjamin_wallace_on_the_price_of_

happiness.html

Ray Anderson

ray_anderson_on_the_business_logi

c_of_sustainability.html

JJ Abrams j_j_abrams_mystery_box.html

Richard Baraniuk

richard_baraniuk_on_open_source_l

earning.html

Dan Barber

dan_barber_s_surprising_foie_gras_

parable.html

Michelle Obama michelle_obama.html

Elizabeth Gilbert elizabeth_gilbert_on_genius.html

Dave Eggers

dave_eggers_makes_his_ted_prize_

wish_once_upon_a_school.html

George Smoot

george_smoot_on_the_design_of_th

e_universe.html

Noah Feldman

noah_feldman_says_politics_and_rel

igion_are_technologies.html

Janine Benyus

janine_benyus_shares_nature_s_desi

gns.html

Majora Carter

majora_carter_s_tale_of_urban_rene

wal.html

Stewart Brand

/stewart_brand_on_squatter_cities.ht

ml

Robert Neuwirth

robert_neuwirth_on_our_shadow_cit

ies.html

Mae Jemison

mae_jemison_on_teaching_arts_and

_sciences_together.html

hal-0

0636

624,

ver

sion

1 -

27 O

ct 2

011

22

Endnotes

Gever Tulley

http://www.ted.com/speakers/gever_

tulley.html

Rob Forbes rob_forbes_on_ways_of_seeing.html

Joseph Pine

joseph_pine_on_what_consumers_w

ant.html

Mike Rowe

mike_rowe_celebrates_dirty_jobs.ht

ml

Deborah Scranton

deborah_scranton_on_her_war_tape

s.html

Jenny 8. Lee

jennifer_8_lee_looks_for_general_ts

o.html

Nancy Etkoff

nancy_etcoff_on_happiness_and_wh

y_we_want_it.html

Jill Bolte Taylor

jill_bolte_taylor_s_powerful_stroke_

of_insight.html

Philip Rosedale the_inspiration_of_second_life.html

John Markoff john_markoff_on_newspapers.html

Penelope Boston penelope_boston.html

Catherine Mohr

catherine_mohr_surgery_s_past_pre

sent_and_robotic_future.html

Sylvia Earle

sylvia_earle_s_ted_prize_wish_to_p

rotect_our_oceans.html

Samantha Power

samantha_power_on_a_complicated

_hero.html

Alisa Miller

alisa_miller_shares_the_news_about

_the_news.html

Jill Tarter

jill_tarter_s_call_to_join_the_seti_se

arch.html

Susan Savage-

Rumbaugh

susan_savage_rumbaugh_on_apes_t

hat_write.html

Mary Roach

mary_roach_10_things_you_didn_t_

know_about_orgasm.html

hal-0

0636

624,

ver

sion

1 -

27 O

ct 2

011

23

i It is unclear which data for American English are taken from Vaux‟s 2002 polling

figures and which data stem from discussions with Dauer or the work of Shitara,

both of which Wells used in preparing the 1999 edition (LPD, 2008, x-xi). ii The Free Dictionary seems to use one man's live, human-being (not Text-To-

Speech) voice to pronounce individual words in American English. Clicking on the

flags gives voices which are definitely TTS, but clicking on the megaphone

symbol next to the word usually gives the voice of one man; there is enough

intonational variation to believe that this is not TTS. The Merriam-Webster on-line

dictionary also has audio pronunciation and they are recorded by real human

beings (e-mail confirmation July 27, 2009). The CD which accompanies the 2008

edition of the LPD has both RP and GAE pronunciations of headwords recorded

by real-human beings but, like several other on-line dictionaries, it does not

provide models of all items under each headword.

hal-0

0636

624,

ver

sion

1 -

27 O

ct 2

011